A Path Towards Clinical Adaptation of Accelerated MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accelerated Magnetic Resonance Imaging (MRI) uses deep learning for faster image reconstruction from sparse data.
• Current deep learning MRI reconstruction methods often lack clinical validation due to simulated environments and resource constraints.

Purpose of the Study:

• To augment neural network MRI reconstructors for enhanced clinical relevance.
• To address challenges in artifact detection, performance variability, multi-anatomy learning, and limited data in accelerated MRI.

Main Methods:

• Developed a Convolutional Neural Network (ConvNet) for detecting MRI image artifacts.
• Trained reconstructors on MR signal data with variable acceleration factors.
• Introduced a novel loss function to prevent catastrophic forgetting in multi-anatomy reconstruction.
• Utilized simulated phantom data for pre-training reconstructors with limited clinical data.

Main Results:

• Achieved a 79.1% F-score for artifact detection using the proposed ConvNet model.
• Demonstrated up to a 2% improvement in average performance during clinical patient scans through variable acceleration factor training.
• Successfully implemented a loss function to enable reconstruction of multiple anatomies and orientations without performance degradation.
• Showcased the efficacy of simulated data pre-training for reconstructors with limited clinical datasets and computational resources.

Conclusions:

• The proposed augmentations enhance the clinical applicability of deep learning-based accelerated MRI reconstruction.
• These methods offer a viable strategy for adapting accelerated MRI to clinical settings, improving efficiency and diagnostic accuracy.